摘要
根据水文时间序列多尺度、非平稳特性,基于“分解-预测-重构”思想,提出小波包分解(WPD)-爬行动物搜索算法(RSA)-极限学习机(ELM)组合多步预测模型,并应用于云南省革雷站月径流、月降水预测。首先介绍RSA原理,选取6个标准函数在不同维度条件下对RSA进行仿真测试,并与哈里斯鹰优化(HHO)、旗鱼优化(SFO)等4种算法的仿真结果进行比较;其次利用WPD对实例水文时序数据进行3层小波包分解,以降低水文序列数据的复杂度;并在延迟时间为1的情况下,采用改进的虚假邻近点法(Cao方法)确定各子序列分量的输入维度;最后通过各分量训练样本构建ELM适应度函数,采用RSA对适应度函数进行寻优,利用寻优获得的最佳ELM输入层权值和隐含层偏值,建立WPD-RSA-ELM模型,对各子序列分量进行超前一步至超前五步预测,将预测结果加和重构得到最终多步预测结果。结果表明:RSA具有较好的寻优精度和全局搜索能力,寻优精度优于HHO、GWO、SFO、PSO算法。WPD-RSA-ELM模型对实例月径流、月降水超前一步至超前五步预测的平均绝对百分比误差分别在0.23%~3.46%和0.60%~9.63%之间,具有较高的预测精度。WPD-RSA-ELM模型预测误差随着预测步数的增加而增大,超前预测步数越多,预测精度越低,预测效果越不理想。
According to the multi-scale and non-stationary characteristics of hydrological time series,based on the idea of“decomposition-prediction-reconstruction”,a wavelet packet decomposition(WPD)-reptile search algorithm(RSA)-extreme learning machine(ELM)combined multi-step forecasting model is proposed.And it is applied to the forecast of monthly runoff and monthly precipitation of Gelei Station in Yunnan Province.First,we introduce the principle of RSA,select 6 standard functions to perform simulation tests on RSA under different dimensional conditions,and compare the simulation results with the simulation results of the Harris Hawk Optimization(HHO)algorithm and the Sailfish Optimization(SFO)algorithm.Second,we use WPD to perform 3-layer wavelet packet decomposition on the example hydrological time series data to reduce the complexity of the hydrological sequence data;and when the delay time is 1,the Cao method is used to determine the input dimensions of each sub-sequence component.Finally,the sample constructs the ELM fitness function through the training of each component,we use RSA to optimize the fitness function,and use the best ELM input layer weights and hidden layer bias values obtained by the optimization to establish a WPD-RSA-ELM model to advance each subsequence component.To the first five-step prediction,the prediction results are added and reconstructed to obtain the final multi-step prediction results.The results show that RSA has better optimization accuracy and global search ability,and the optimization accuracy is better than HHO,GWO,SFO,and PSO algorithms.The WPD-RSA-ELM model predicts the average absolute percentage errors of monthly runoff and monthly precipitation from one step ahead to five steps ahead,respectively,between 0.23%~3.46%and 0.60%~9.63%,which has high prediction accuracy.The prediction error of WPD-RSA-ELM model increases as the number of prediction steps increases.The more advanced prediction steps,the lower the prediction accuracy and the less ideal the prediction effect.
作者
李新华
崔东文
LI Xinhua;CUI Dongwen(Yunnan Xingdian Group Co.,Ltd.,Wenshan 663000,Yunnan,China;Yunnan Province Wenshan Water Bureau,Wenshan 663000,Yunnan,China)
出处
《水利水电技术（中英文）》
北大核心
2022年第11期69-77,共9页
Water Resources and Hydropower Engineering
基金
国家自然科学基金项目“澜沧江非一致性径流演变规律及驱动机制研究”(91547205)。
关键词
水文预测
小波包分解
爬行动物搜索算法
极限学习机
仿真测试
多步预测
hydrological forecast
wavelet packet decomposition
reptile search algorithm
extreme learning machine
simulation test
multi-step forecast
